We adapted a previously-developed model of HIV and reproductive health in South Africa (MicroCOSM, or Microsimulation for the Control of South African Morbidity and Mortality). This is an agent-based model, which has previously been used to evaluate structural drivers of HIV and STIs [58–60]. Briefly, the model simulates a nationally-representative sample of the South African population, with each individual (‘agent’) being randomly assigned a set of characteristics: demographic (age, sex, and race), socio-economic (education, urban/rural location, migration, and incarceration history), psychological (gender norms, conscientiousness), healthcare access (use of hormonal contraception, HIV prevention, and HIV testing/treatment), and behavioural (alcohol consumption, propensity for concurrent partners, sexual experience, sexual preference, and current relationship/marital status). Individuals are classified as ‘high-risk’ or ‘low-risk’ based on their propensity for concurrent partners, and four types of sexual relationship are modelled (Fig 1A): long-term marital/cohabiting, short-term noncohabiting, casual/once-off (characterised as ‘transactional sex’) and commercial sex (between sex workers and clients). When new partnerships are formed, individuals are linked to other individuals in the simulated population, and the transmission of HIV/STIs is modelled based on specified probabilities of transmission per condomless sex act (with multiplier adjustments to account for factors such as the HIV–positive partner’s HIV viral load and male circumcision). A woman’s monthly probability of conception is similarly based on the woman’s numbers of sexual partners, age, and contraceptive use. The simulation begins in 1985, with HIV being introduced into the simulated population in 1990.

Socio-economic status has a number of dimensions, which can influence sexual behaviour in different ways, and socio-economic status is often measured in terms of household wealth/income [61]. We therefore chose to model socio-economic status based on a combination of household- and individual-level characteristics. We extended the model to identify family links between individuals (parents, children, and siblings), and we grouped individuals into households based on family and marital relationships. We modelled household formation and dissolution dynamically, calibrating the model to national survey data on household composition (see section 1.2 of S1 Text).

The modelling of educational attainment has been described previously [59], and is illustrated in Fig 1B. We further extended the model to simulate employment in individuals aged 15–64 who are not currently in school/studying or incarcerated. Rates of entering employment are assumed to depend on age, sex, educational attainment, race, urban/rural location, and parity (in women). In addition, we assumed people living with HIV experience a reduced odds of employment if they are untreated and have low CD4 counts [62–64]. We calibrated the model to national survey data on employment levels by age, sex, and race (see section 1.1 of S1 Text).

We modelled four sources of household income in the baseline scenario: salaries/wages (for each employed member in the household), child support grant payments, the state old age pension, and private pensions (Fig 1C). A more detailed description of these four income sources is provided in section 1.3 of S1 Text. The adjusted per capita household income (APCHI) is calculated as

where T is the total household income, nA and nC are the numbers of adults (15+) and children respectively in the household, α is the relative cost of supporting children (relative to adults) and θ represents the economies of scale in meeting the needs of larger households. In line with previous South African studies [65], we set α = 0.5 and θ = 0.9; other equivalence scales are reviewed elsewhere [66].

Fig 1D illustrates the hypothesised causal pathways linking socio-economic status and HIV/reproductive health outcomes. We assumed economic strengthening interventions affect socio-economic variables (educational attainment, employment, and household income), and these in turn affect sexual behaviour; sexual behaviour then influences HIV, STIs, and pregnancy (which can in turn influence socio-economic status, e.g., pregnancy causing school dropout). Table 1 summarises the key parameters that relate socio-economic status to sexual risk behaviour and health seeking behaviour (described more fully in sections 1.4 and 1.5 of S1 Text). We modelled three intervention types, all targeting households with adjusted per capita income below the national average: (1) school support to reduce school dropout (the intervention is also assumed to increase the chance of re-enrolment in youth aged 13–22 who have dropped out of school); (2) vocational training (for ages 20–49); and (3) unconditional cash transfers. We selected these interventions from a more complete list of economic strengthening interventions [45], based on the availability of data regarding their impacts on HIV/STI/teenage pregnancy outcomes. The school support intervention is assumed to include both material support (e.g., school uniforms, cost of school transport) and nonmaterial support (e.g., attendance monitoring and counselling), both of which reduce the probability of school dropout. We modelled the effects of the two components separately, with half of the cash equivalent value of the material support being added to the household income, in the same way as for cash transfers. We assumed vocational training increases the odds of finding employment in currently unemployed individuals. Cash transfer amounts are added to total household income as another source of income.

We adopted a Bayesian approach to model calibration [91], similar to that in a previous application of our model [58]. Prior distributions were specified to represent the uncertainty in key model parameters (Table 1). We specified hurdle distributions to represent the uncertainty in the effect of socio-economic status on sexual risk behaviour, with the hurdle representing a nonzero probability of a null relationship (this was included to avoid excessive prior confidence in the purported causal relationship). We set the nonzero probabilities based on a review of the strength of evidence from RCTs; a more detailed description of the prior distributions is provided in section 2.1 of S1 Text.

We calibrated the model to the results of RCTs of economic strengthening interventions conducted in sub-Saharan Africa. RCTs were identified from recent reviews of economic strengthening interventions [45,46,48,49]. We classified included trials as being either ‘pure’ cash transfer interventions (including both conditional and unconditional cash transfers, but without any strong conditioning on school attendance), school support interventions (which typically aimed to promote school retention, often through the provision of financial support, or through cash transfers conditional on school attendance) and vocational training programmes (directed to individuals who were out of school, providing training to improve their employment prospects and/or credit to enable them to establish their own business). We excluded less frequently studied interventions (e.g., savings programmes, food assistance, and financial education) due to limited RCT data available on their effects [45]. Intervention effects were calculated on a natural log scale, with associated standard errors (Table ZB in S1 Text).

A sample of 5,000 parameter combinations was randomly drawn from the prior distributions in Table 1. For each parameter combination, we ran the model eight times: twice in the ‘baseline’ scenario, twice in the cash transfer scenario, twice in the school support scenario, and twice in the vocational training scenario (two simulations being necessary in order to quantify the extent of stochastic variation in model outputs; the average of the two results was calculated for the purpose of estimating intervention effects). For the purpose of calibration to historic RCT data, we assumed all interventions started in 2005 (close to the average start year of the included RCTs), and we set the annual value of the cash transfer in 2005 to R800 (equivalent to $117 in 2005), the average value of the cash transfers in the included RCTs. We assumed the interventions were directed to households with log APCHI below the national average (8.68 in 2005, equivalent to R5856 or $861 annually). The modelled intervention effect was calculated as the log of the odds ratio/relative risk for the outcome of interest (comparing the intervention and baseline scenarios). We calculated a likelihood value by comparing the modelled intervention effect and the reported intervention effect, assuming the log difference followed a normal distribution with zero mean. Log likelihood values were summed for all outcomes, across all RCTs, to calculate a total likelihood for each parameter combination; separate likelihoods were also calculated for each of the three intervention types.

We drew a posterior sample of 100 parameter combinations from the initial set of 5,000 parameter combinations, using the likelihood values as weights. This sample was used to calculate the posterior means and 95% confidence intervals (CIs), running the model five times for each parameter combination to reduce stochastic variation.

We validated the model by comparing the modelled odds ratios for the associations between socio-economic status and HIV/sexual risk behaviour, as observed in four nationally-representative household surveys in 2005, 2012, 2016, and 2017 [79,92].

In all scenarios, we set STI transmission probabilities at the median of the distribution of best-fitting parameters identified in previous model calibrations [93]. We adjusted HIV transmission probabilities per sex act to yield plausible estimates of HIV incidence and prevalence (see section 3.2 of S1 Text). We estimated teenage fertility rates, per year of sex with a single partner in the absence of contraception, from age- and race-specific fertility rates in the baseline scenario [94].

We calculated the proportion of incident HIV cases/STIs/teenage births that were attributable to low socio-economic status, over the 2000–2020 period, by running a counterfactual scenario in which the sexual behaviour and health seeking behaviour of all individuals are the same as might be expected (a) if they had completed tertiary education, (b) if they remained in education at least to age 21, (c) if they were employed, and (d) if their adjusted per capita household income were no lower than the national mean in the baseline scenario (with the changes in behaviour occurring from mid-2000). We calculated the population attributable fraction (PAF) as the proportionate difference in cumulative HIV cases/STIs/teenage births, over the 2000–2020 period, between this counterfactual scenario and the baseline scenario. Partial rank correlation coefficients (PRCCs) were calculated to assess associations between PAFs and each of the parameters that were changed in the counterfactual scenario [95].

We modelled three future intervention scenarios: school support, vocational training, and cash transfers. Future intervention impacts were considered over the period from 2025 to 2040. We assumed interventions were limited to households with adjusted per capita household income below the national average in the baseline scenario. We set the annual value of the cash transfer to $227 per household in the case of cash transfer interventions and $114 per eligible youth in the case of school support (at 2023 exchange rates), increasing in line with inflation [96].

All model code and data inputs are available from https://github.com/leighjohnson/MicroCOSM.

Our model was in reasonable agreement with observed HIV prevalence and incidence trends in South Africa, external estimates of the incidence of gonorrhoea and chlamydia [97] and data on teenage fertility (Fig 2A–2D). The model was also in agreement with census and survey data showing rising levels of high school completion among youth, and stable employment levels (Fig 2E and 2F). The model also estimated rising income together with declining income inequality, the result of increasing levels of social welfare expenditure, and both trends were validated by estimates from Statistics South Africa (Fig 2G and 2H).

Posterior estimates of model parameters were largely consistent with prior distributions, although intervention effect parameters differed significantly: vocational training was estimated to be less effective in reducing unemployment (posterior odds ratio [OR] 0.93, prior OR 0.87), while posterior estimates of school support impacts were mixed, with material support being more effective, compared to prior assumptions about effects on school dropout (Table ZD in S1 Text). The posterior estimates of RCT effects were in good agreement with the data, although the model was unable to match the large reductions in HIV and herpes prevalence and pregnancy incidence in one trial [69] (Figs N–P in S1 Text). The model was also consistent with most of the validation data, although the model under-estimated the strength of the negative association between HIV prevalence and education in women (Fig Q in S1 Text).

Low socio-economic status accounted for 13% of new HIV infections, 7% of incident STIs (gonorrhoea, chlamydia, and trichomoniasis) and 31% of teenage births in South Africa, over 2000−2020 (Table 2). However, because of uncertainties regarding effect sizes, CIs around these PAFs were wide (1,50% for HIV; −1,19% for STIs; and 10,76% for teenage births). PAF estimates were similar for men and women. Table 3 shows the socio-economic effect parameters that account for the most uncertainty. The effect of education on condom use was the most significant correlate of the HIV PAF (PRCC = 0.87) and the STI PAF (PRCC = 0.90), but was less significant as a correlate of the teenage birth PAF (PRCC = 0.57). The relative rate of female sexual debut while in school was the most significant correlate of the teenage birth PAF (PRCC = −0.95) and was also significantly correlated with the HIV and STI PAFs; the relative rate of male sexual debut while in school was also weakly associated with the teenage birth PAF. The effect of male employment on partnership formation was positively associated with all the PAFs, although this was only of borderline significance. Other parameters were not statistically significant, and fewer parameters were significant when standard correlation coefficients were calculated in place of PRCCs (Table ZG in S1 Text).

In the absence of any economic strengthening interventions, our model suggests there would be little change in income inequality and high school completion over the next 15 years, although employment is expected to increase modestly as a result of population ageing (Fig 3). The cash transfer intervention would reduce income inequality, with the Palma ratio (the ratio of income in the top decile to that in the four lowest deciles) decreasing from 8.5 to 7.8 by 2030; school support and vocational training would have more modest effects on income inequality (Fig 3A). The school support intervention would significantly increase the proportion of South African youth who have completed high school, from 30% to 38% by 2040, but the other interventions would have no effects on school completion (Fig 3B). Vocational training interventions would only slightly increase employment levels, while school support would slightly reduce the fraction of the population employed in the short-term, due to young adults remaining in school for longer (Fig 3C).

When considering intermediate behavioural outcomes, the effects of the cash transfer and vocational training interventions are close to zero (Table 4). However, the school support intervention is expected to significantly reduce the proportion of sexually active young females who are not using hormonal contraception, the proportion of adults in concurrent partnerships, and the proportion of men engaging in casual sex (the latter two effects being due to assumed effects of education on inequitable gender norms, and assumed effects of inequitable gender norms on male risk behaviour [58]). The school support intervention may also reduce the proportion of adolescents who are sexually experienced and the number of early marriages, although these reductions are not statistically significant (Table 4).

Our model suggests that economic strengthening interventions would have no significant impact on most reproductive health outcomes (Table 4). Over 2025–2040, schooling support is the intervention that would have the greatest impact, with an anticipated 5% reduction in teenage births (which is strongly correlated with the relative rate of sexual debut for girls who remain in school [PRCC = −0.84]) and a 1% reduction in HIV and STI incidence (which is strongly correlated with the relative rate of sexual debut among girls in school [PRCC = −0.58 for STIs and −0.27 for HIV] and the increase in condom use per year of education [PRCC = 0.58 for STIs and 0.25 for HIV]). For all other intervention/outcome combinations, the expected impact is 0%, with lower CI limits below zero (partly a reflection of residual stochastic variation). Vocational training is unlikely to have a significant impact if it is limited to females (for example, a 0% reduction [95% CI: −1,1] in STIs). Intervention impacts are similar in males and females. Proportionate reductions in HIV, STIs and teenage pregnancies are also similar over the 2025–2040 period, i.e., there is no suggestion of the intervention impact waning or improving at longer durations (Table ZH in S1 Text).